Sie sind auf Seite 1von 9

Environmental Pollution 233 (2018) 1086e1094

Contents lists available at ScienceDirect

Environmental Pollution
journal homepage:

Estimating spatiotemporal distribution of PM1 concentrations in China

with satellite remote sensing, meteorology, and land use information*
Gongbo Chen a, Luke D. Knibbs b, Wenyi Zhang c, Shanshan Li a, Wei Cao d, Jianping Guo e,
Hongyan Ren d, Boguang Wang f, Hao Wang g, Gail Williams b, N.A.S. Hamm h,
Yuming Guo a, *
Department of Epidemiology and Preventive Medicine, School of Public Health and Preventive Medicine, Monash University, Melbourne, Australia
School of Public Health, The University of Queensland, Brisbane, Australia
Center for Disease Surveillance & Research, Institute of Disease Control and Prevention, Academy of Military Medical Science, Beijing, China
Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences, Beijing, China
Sate Key Laboratory of Severe Weather, Chinese Academy of Meteorological Sciences, Beijing, China
Institute for Environmental and Climate Research, Jinan University, Guangzhou, China
Air Quality Studies, Department of Civil and Environmental Engineering, Hong Kong Polytechnic University, Hong Kong, China
Faculty of Geo-Information Science and Earth Observation (ITC), University of Twente, Enschede, The Netherlands

a r t i c l e i n f o a b s t r a c t

Article history: Background: PM1 might be more hazardous than PM2.5 (particulate matter with an aerodynamic
Received 29 July 2017 diameter  1 mm and 2.5 mm, respectively). However, studies on PM1 concentrations and its health
Received in revised form effects are limited due to a lack of PM1 monitoring data.
19 September 2017
Objectives: To estimate spatial and temporal variations of PM1 concentrations in China during 2005
Accepted 4 October 2017
e2014 using satellite remote sensing, meteorology, and land use information.
Available online 13 October 2017
Methods: Two types of Moderate Resolution Imaging Spectroradiometer (MODIS) Collection 6 aerosol
optical depth (AOD) data, Dark Target (DT) and Deep Blue (DB), were combined. Generalised additive
model (GAM) was developed to link ground-monitored PM1 data with AOD data and other spatial and
Aerosol optical depth temporal predictors (e.g., urban cover, forest cover and calendar month). A 10-fold cross-validation was
Meteorology performed to assess the predictive ability.
Land use Results: The results of 10-fold cross-validation showed R2 and Root Mean Squared Error (RMSE) for
China monthly prediction were 71% and 13.0 mg/m3, respectively. For seasonal prediction, the R2 and RMSE
were 77% and 11.4 mg/m3, respectively. The predicted annual mean concentration of PM1 across China
was 26.9 mg/m3. The PM1 level was highest in winter while lowest in summer. Generally, the PM1 levels
in entire China did not substantially change during the past decade. Regarding local heavy polluted
regions, PM1 levels increased substantially in the South-Western Hebei and Beijing-Tianjin region.
Conclusions: GAM with satellite-retrieved AOD, meteorology, and land use information has high pre-
dictive ability to estimate ground-level PM1. Ambient PM1 reached high levels in China during the past
decade. The estimated results can be applied to evaluate the health effects of PM1.
© 2017 Elsevier Ltd. All rights reserved.

1. Introduction problems causing 1.6 million deaths nationwide annually (Kan

et al., 2009; Rohde and Muller, 2015). Fine particulate matter
With the rapid growth of the economy and expansion of the with aerodynamic diameter 2.5 mm (PM2.5) has attracted
urban population, China is experiencing serious air pollution increasing public concern and its adverse health effects have been
documented by numerous studies (Cao et al., 2012; Dockery and
Stone, 2007; Ma et al., 2011; Yang et al., 2012). Particulate matter
This paper has been recommended for acceptance by B. Nowack. with aerodynamic diameter 1 mm (PM1), a major part of fine
* Corresponding author. Department of Epidemiology and Preventive Medicine, particulate matter mass, has seldom been studied e either to
School of Public Health and Preventive Medicine, Monash University. Level 2, 553 St
investigate its spatiotemporal variation or to investigate its asso-
Kilda Road, Melbourne, VIC 3004, Australia.
E-mail address: (Y. Guo). ciations with health outcomes. PM1 accounts for more than 80% of
0269-7491/© 2017 Elsevier Ltd. All rights reserved.
G. Chen et al. / Environmental Pollution 233 (2018) 1086e1094 1087

ambient PM2.5 mass at some locations, particularly in China of DT AOD and DB AOD was available via NASA (MODIS Aerosols
(Cabada et al., 2004; Li et al., 2015; Wang et al., 2015). Due to its Merged Dark Target Deep Blue Product) in which DB AOD values
smaller particle size, PM1 might be more harmful than PM2.5 and were discarded with Normalized Difference Vegetation Index
more strongly associated with some health outcomes (Chen et al., (NDVI) values greater than 0.3 (Levy et al., 2013). To increase the
2017; Lin et al., 2016). spatial coverage of AOD data, this merged product was not used in
To fill in spatial and temporal gaps left by ground-based mea- this study, we instead, obtained DT and DB AOD products separately
surements of air pollution, satellite remote sensing has been used and used the approach of Ma et al. (2015) to fill the missing values
successfully in recent years to predict concentrations of air pollu- in both products and combine them (Ma et al., 2015). A model
tion at locations with sparse ground monitoring data, based on the linking DT and DB AOD data was developed to fill the missing
validated relationships between satellite remote sensing and values. This model was based on simple linear regression:
ground measurements (Hu et al., 2014b; Just et al., 2015; Kloog
et al., 2012; Koelemeijer et al., 2006). Aerosol optical depth AODDT ¼ a þ b *AODDB þ ε
(AOD), also referred to as aerosol optical thickness (AOT), is the
most widely used satellite-retrieved atmospheric product that has where: AODDT is the DT AOD value; AODDB is the DB AOD value; a is
been used to predict air pollution concentrations. AOD is a measure the intercept and b is the slope coefficient and ε is a normally
of the attenuation of solar radiation by aerosols in the atmosphere distributed residual. This model was used to fill the missing values
and is correlated with PM concentration at ground level in many of DT AOD when values of DB AOD were valid, and vice-versa.
regions (Koelemeijer et al., 2006; Lee et al., 2011). Previous studies Ground-measured AOD data in China during the study period
have reported satellite-retrieved concentrations of PM2.5 and PM10 were downloaded from the Aerosol Robotic Network (AERONET)
(particulate matter with aerodynamic diameter  10 mm) in China ( to combine DT and DB AOD data.
using AOD and other predictors with high spatial resolution and The AERONET AOD data at 675 nm and 440 nm were extracted to
predictive ability (Fang et al., 2016; Ma et al., 2015; Meng et al., interpolate the AOD values at 550 nm, which were then linked with
2015; Zheng et al., 2016), but no study has reported satellite- DT and DB AOD by location and time (Jing-Mei et al., 2010; Sayer
retrieved concentrations of PM1. et al., 2013). Details about the interpolation are shown in “Inter-
In this study, we aimed to combine daily ground monitoring polation of AOD at 550 nm” in the Supplemental Material. The
data of PM1 and Moderate Resolution Imaging Spectroradiometer differences between DB AOD (or DT AOD) and AERONET AOD were
(MODIS) Collection 6 AOD data with other spatial and temporal calculated, and the inverse variances of these differences were used
predictors to estimate the concentrations of PM1 across China as weight to combine DB and DT AOD data. The locations of 40
during 2005e2014. AERONET sites providing ground-measured AOD data are shown in
Fig. S1 in the Supplemental Material. Compared with merged AOD
2. Materials product available at Aqua MODIS C6, the combined AOD data
derived using methods above showed a substantial improvement
2.1. Ground measurements of PM1 and PM2.5 in spatial coverage (Ma et al., 2015).

Hourly ground-level measurements of PM1 and PM2.5 were 2.3. Meteorological data
obtained from 77 stations of China Atmosphere Watch Network
(CAWNET) from November 2013 to July 2014 and September 2013 Daily meteorological data were obtained from 824 weather
to December 2014, respectively (Guo et al., 2009, 2017). We used stations of the China meteorological data sharing service system
this time span because it had contemporaneous measurements of during 2005e2014 ( The locations of these
both particle sizes. Hourly concentrations of PM1 and PM2.5 during weather stations are shown in Fig. S2 in the Supplemental Material.
the study period were measured with the GRIMM 180 (Grimm 180 Daily mean temperature, relative humidity, barometric pressure
Multi-channel Aerosol Spectrometer) environmental dust monitors and wind speed were used in this study. Meteorological variables in
(Grimm and Eatough, 2009). This instrument is an optical particle areas not covered by weather stations were interpolated using
counter (OPC) with 31 size channels and operates at a flow rate of “Micro krig” in the R package “fields” (Furrer et al., 2009). The de-
1.2 L/min. The recorded particle size number distribution between tails of this interpolation are shown in “Interpolation of meteoro-
0.25 mm and 32 mm is then calibrated to a particle mass concen- logical variables” in the Supplemental Material.
tration. Details about the measurements and calibration were re-
ported previously (Wang et al., 2015). Daily mean concentrations of 2.4. Land use information and vegetation data
PM1 and PM2.5 were calculated as (Cdaily ¼ 24 1 Chour =24), where C
denotes the PM1 or PM2.5 concentrations. Two approaches were Annual land cover data (including urban cover, forest cover, and
applied to control the quality of PM1 measurements (Guo et al., water cover) from 2004 to 2012 at a spatial resolution of 500 m
2009). The locations of 77 monitoring stations are shown in Fig. 1. were obtained from Global Mosaics of the standard MODIS land
More stations were located in Eastern and Central China, especially cover type data Collection 5.1 product of Global Land Cover Facility
for South-Eastern coastal areas, than Western China. ( (Friedl et al., 2010). Land cover data in 2012
were used for prediction of study years 2012e2014, as the data
2.2. Aerosol optical depth were not available during 2013 and 2014. In total, there are 17 types
of land cover variables in the satellite data sets and the pixel size is
Two types of daily MODIS AOD data, Dark Target (DT) and Deep 500 m. The percentages of forest cover (or other types of land
Blue (DB), from the Aqua Atmosphere Level 2 Product Collection cover) were calculated by dividing the count of forest cover pixels
6 at 10-km resolution and covering China were downloaded from by the count of pixels for all types of land cover within a given
NASA Level-1 and Atmosphere Archive & Distribution System radius buffer. MODIS Level 3 monthly average NDVI products with a
Distributed Active Archive Centre for 2005 to 2014 (https:// spatial resolution of 0.1 (z10 km) during the study period were DT and DB AOD downloaded from the NASA Earth Observatory (http://neo.sci.gsfc.
were then combined with an Inverse Variance Weighting Method Further information about these data products has been
after filling the their gaps (Ma et al., 2015). A merged AOD product previously described (Hamm et al., 2015b).
1088 G. Chen et al. / Environmental Pollution 233 (2018) 1086e1094

Fig. 1. Locations of 77 stations with ground-based measurements of PM1 and PM2.5.

2.5. Other spatial predictors 2014. Based on the high correlation between PM1 and PM2.5 con-
centrations and their relationships with temperature and relative
Aqua and Terra active fires during the study period were humidity (Lee et al., 2006b; Li et al., 2015; Wang et al., 2015), daily
downloaded from NASA Fire Information for Resource Manage- concentrations of PM1 at the 77 stations during the periods from
ment System ( Sep 2013 to Nov 2013 and from Jul 2014 to Dec 2014 were inter-
data/firms) (Hu et al., 2014a). Daily counts of fire spots within a polated with the following generalised additive model (GAM):
buffer of 75 km were calculated for each ground monitoring station
and grid cell created. The global Shuttle Radar Topography Mission PM1 ¼ s(PM2.5) þ s(TEMP)þ s(RH)
(SRTM) Digital Elevation Model (DEM) has a resolution of 3 arc-
seconds (approximately 90 m) (Hamm et al., 2015b). SRTM where: TEMP and RH refer to daily mean temperature and relative
version 4 elevation data for China were downloaded from The humidity, respectively. The degrees of freedom for smooth terms
CGIAR Consortium for Spatial Information (http://srtm.csi.cgiar. were automatically selected by GAM. This interpolation was per-
org/). The elevation for each monitoring station was extracted formed for each of 77 stations separately and together it captured a
and mean value of all elevation pixels fell in each grid cell was large proportion of variability in PM1 (R2 ¼ 93%). The interpolated
calculated. PM1 data covering 15 months were more suitable than the original
9-month ground monitoring data to capture temporal trends and
seasonality of PM1 concentrations.
2.6. Model development and validation AOD, meteorological variables, and elevation were determined
at each measurement point, while land use variables were deter-
A 0.1-degree grid (z10 km) covering China with 96 104 grid mined at a range of buffers from 100 m to 10 km (Knibbs et al.,
cells was created to integrate spatial and temporal predictors and 2014). The total number of active fires within 75 km of each site
develop models. Daily values of meteorological variables were was counted (Hu et al., 2014c).
interpolated for each grid cell based on daily meteorological data Our approach to model development was informed by recent
from 824 weather stations. For predictors with a resolution of 0.1 PM2.5 studies in China that described predictor variables were also
(e.g., AOD, meteorological variables and NDVI), values were directly potentially associated with PM1 (Fang et al., 2016; Ma et al., 2014,
extracted for each grid cells. For predictors with a resolution higher 2015). We used a GAM and began by including AOD and then
than 0.1 (e.g., land cover and elevation), mean values of all pixels of incrementally included other predictors until we reached a parsi-
predictors in each grid cell were calculated and used for prediction. monious model that explained the most variability in PM1. For land
Monthly and annual predictors (e.g., NDVI, urban cover and forest use variables calculated at different buffers that were associated
cover) were linked with daily predictors according to the month or with PM1, we included the buffer distance that offered greatest
year they were collected, as the values of those variables are un- ability to explain PM1. We arrived at the following GAM as being the
likely to change within one month or year, respectively. All spatial best model for predicting daily concentrations of PM1:
and temporal predictors were integrated into the grid for each grid
cell by location (longitude and latitude of the centroid) and cal- PM1 ¼ AODc  province þ s(TEMP)  province þ s(RH)  province
endar date. In this study, the ground monitoring data of PM1 was þ s(WS)  province þ s(BP) þ firesmoke  province þ NDVI
only available for 9 months from November 2013 to July 2014, and  province þ Forest_cover þ Urban_cover
PM2.5 data covered 15 months from September 2013 to December
G. Chen et al. / Environmental Pollution 233 (2018) 1086e1094 1089

þ Water_areas þ month þ Dayofweek þ log(elev) Sensitivity analyses showed that our results were robust; adding
hours of sunshine and population density did not improve the
where: AODc is the combined AOD; province is the province where model performance, and calendar month is more suitable than day
the station was located and it is an interaction term to account for of year or season to account for the long-term trend in PM1. Results
the regional variations of PM1-AOD association; TEMP is daily mean of the sensitivity analyses are shown in Table S8 in the Supple-
temperature ( C); RH is daily mean relative humidity (%); WS is mental Material.
daily mean wind speed (km/h); BP is daily mean barometric pres- Fig. 4 shows the estimated mean concentrations of PM1 during
sure (kPa); firesmoke is the count of fire smoke spots; NDVI is the 2005 through 2014. The mean concentration of PM1 predicted
monthly average NDVI value; Forest_cover is the percentage of across China was 26.9 mg/m3. The highest levels of PM1 (70 mg/m3)
forest cover (3-km radius buffer); Urban_cover is the percentage of were predicted in South-Western Hebei, Beijing and Tianjin. Rela-
urban cover (10-km radius buffer); Water_areas is the percentage of tively high levels of PM1 (50 mg/m3 and <70 mg/m3) were present
water areas (10-km radius buffer); log(elev) is the log transformed in Sichuan, Chongqing, Henan and Liaoning. The lowest levels of
elevation (m). The degrees of freedom for smoothed terms were PM1 (<20 mg/m3) were shown in South-Western and Northern
automatically selected by GAM. remote areas of China including Xizang, Yunnan and Northern In-
Although our dependent variable was daily PM1 in the period ner Mongolia.
from September 2013 to December 2014, we also wanted to Fig. 5 shows the estimated seasonal concentrations of PM1
demonstrate the feasibility of longer-term (i.e., decadal) estimation across China with the highest levels predicted in winter
of PM1. We thus used our final GAM to predict daily concentrations (mean ¼ 45.3 mg/m3) and the lowest in summer (mean ¼ 15.7 mg/
of PM1 for each grid cell created from 2005 to 2014 by capitalizing m3). The levels of PM1 we estimated were similar in spring and in
on historical predictor data, including AOD and meteorological autumn (26.4 mg/m3 and 25.9 mg/m3, respectively). Areas in North-
observations. We also averaged our daily predictions to obtain Eastern China and South-Western Hebei exhibited a substantial
monthly and seasonal estimates of PM1. increase from summer to winter.
To assess the validity of our predictions, a 10-fold cross- Fig. 6 shows the 10-year trends (2005e2014) in PM1 concen-
validation (CV) process was performed using data from 48 days trations estimated in both heavily polluted regions and across the
(of the 478 days total) randomly selected as test set and the rest of entire country. The estimated PM1 levels in China as a whole
the data as the training set. This process was repeated 500 times. exhibited slight increases, with an increase of 2.1 mg/m3 from 2005
The overall adjusted R2 and Root Mean Square Error (RMSE) were to 2014. Modest changes of PM1 levels were observed in Guang-
calculated. Sensitivity analyses were also performed to test the dong, Yangtze River Delta and North-Eastern China. Increased
model's robustness. For example, we added daily hours of sunshine trends of PM1 during the past decade were present in South-
and population density in the final model to check whether they Western Hebei (increased by 8.9 mg/m3), Beijing and Tianjin
improved model performance. We also included other temporal (increased by 8.6 mg/m3) and Chongqing and Eastern Sichuan
predictions (e.g., day of year, season) in the model replacing cal- (increased by 6.5 mg/m3). Locations of these heavy polluted regions
endar month. are shown in Fig. S3 in the Supplementary Material.

3. Results 4. Discussion

In total, 32,675 records of ground-monitored PM1 data from Despite China's well-publicized air pollution problems, studies
September 2013 through December 2014 were included in the on the long-term effects of fine particulate matter on health are
model development. The mean concentration of ground-measured limited due to the lack of ground-level monitoring data, especially
PM1 was 39.1 mg/m3. The lowest level of PM1 was observed in prior to 2013. Statistical models using satellite-retrieved AOD have
Shangri-La (5.0 mg/m3), Yunnan Province, while the highest was in the potential to estimate historical and current exposures to par-
Shijiazhuang (82.1 mg/m3), Heibei Province. Summaries of ground- ticulate matter with good accuracy and spatial resolution by
based measurements of PM1 and weather conditions at the 77 PM1 exploiting the relatively strong relationship between PM2.5 and
monitoring stations are shown in Tables S2eS6 in the Supple- AOD over China, as demonstrated by previous studies (Lee et al.,
mental Material. 2011; Wang and Christopher, 2003; Zhang et al., 2009). To the
Table 1 shows the improvement in performance of the best daily best of our knowledge, this is the first study to estimate PM1 in
GAM with the addition of each successive predictor. The daily AOD- China using satellite remote sensing. Using a combination of MODIS
only model for PM1 showed an R2 of 24%. In other models, mete- AOD data and other spatiotemporal predictors, we estimated daily,
orological variables, especially temperature and relative humidity, monthly and seasonal levels of PM1 from 2005 through 2014 at a
made the greatest contribution to the model performance. The resolution of 0.1 across China. We captured 59%, 71%, and 77% of
model with AOD, and meteorological predictors (temperature and variability in daily, seasonal and monthly PM1 during 2013e14,
relative humidity) had an R2 of 40%. Fig. 2 shows the performance which we then applied to estimate historical levels during the
of final model for predicting PM1 concentrations (step 13 in preceding decade.
Table 1), which explained 58% of the variability in daily PM1 Although only 77 ground monitoring sites were included in this
(RMSE ¼ 21.7 mg/m3). The 10-fold cross-validation showed modest study because PM1 is not routinely monitored, the results of cross-
prediction errors with little bias (Fig. 3). The CV R2 for daily esti- validation indicated the predictive ability of our model is compa-
mation was 59% (RMSE ¼ 22.5 mg/m3, slope ¼ 1.01). rable to that reported in previous study of PM2.5 in China based on
Daily concentrations of PM1 were estimated and the results much larger set of ground monitoring data (Ma et al., 2015). Studies
were averaged to monthly and seasonal mean concentrations. have demonstrated satellite-retrieved AOD is strongly linked with
Monthly and seasonal estimations were improved (R2 ¼ 0.74 and particles between 0.1 and 2.0 mm (Diner et al., 1998; Kahn et al.,
0.82, respectively, RMSE ¼ 12.0 mg/m3 and 9.0 mg/m3, respectively) 2001), and the particle size of PM1 is right within that range.
(Fig. 2). The 10-fold cross-validation shows that higher predictive Additionally, ground monitoring data indicated ambient PM1
ability was observed for monthly estimation (R2 ¼ 71%, accounted for 66%e91% of PM2.5 in China, and with PM1 and
RMSE ¼ 13.0 mg/m3, slope ¼ 0.96) and seasonal estimation meteorological factors, most variations of PM2.5 concentrations can
(R2 ¼ 77%, RMSE ¼ 11.4 mg/m3, slope ¼ 1.02) (Fig. 3). be explained (Lee et al., 2006a; Wang et al., 2015).
1090 G. Chen et al. / Environmental Pollution 233 (2018) 1086e1094

Table 1
Steps for selecting the best model for predicting daily PM1.

Step Variable in model R2 GCV


1 AOD* 25% 914.3

2 AOD*þTemperature* 36% 778.8
3 AOD*þTemperature*þRelative humidity* 40% 731.2
4 AOD*þTemperature*þRelative humidity*þWind speed* 43% 697.0
5 AOD*þTemperature*þRelative humidity*þWind speed*þBarometric pressure 43% 696.9
6 AOD*þTemperature*þRelative humidity*þWind speed*þBarometric pressure þ Month 49% 628.8
7 AOD*þTemperature*þRelative humidity*þWind speed*þBarometric pressure þ Month þ Forest Cover 49% 623.6
8 AOD*þTemperature*þRelative humidity*þWind speed*þBarometric pressure þ Month þ Forest Cover þ Urban Cover 49% 622.4
9 AOD*þTemperature*þRelative humidity*þWind speed*þBarometric pressure þ Month þ Forest Cover þ Urban Cover þ Water Cover 49% 622.3
10 AOD*þTemperature*þRelative humidity*þWind speed*þBarometric pressure þ Month þ Forest Cover þ Urban Cover þ Water Cover þ Fire smoke* 50% 619.2
11 AOD*þTemperature*þRelative humidity*þWind speed*þBarometric pressure þ Month þ Forest Cover þ Urban Cover þ Water Cover þ Fire 52% 587.4
12 AOD*þTemperature*þRelative humidity*þWind speed*þBarometric pressure þ Month þ Forest Cover þ Urban Cover þ Water Cover þ Fire 53% 585.0
smoke*þNDVI*þDay of week
13 AOD*þTemperature*þRelative humidity*þWind speed*þBarometric pressure þ Month þ Forest Cover þ Urban Cover þ Water Cover þ Fire 58% 522.5
smoke*þNDVI*þDay of week þ Elevation

*Refers to variables with “province” as an interaction term in the model.

Fig. 2. Scatterplots of model fitting for daily, monthly and seasonal estimation of PM1 concentrations (mg/m3).

Fig. 3. Scatterplots of 10-fold cross-validation for daily, monthly and seasonal estimation of PM1 concentrations (mg/m3).

The overall temporal trends and seasonality of PM1 in China in in some heavily-polluted regions were relatively high but not the
our study were also consistent with previous studies on PM2.5, highest observed in our study including the Yangtze River Delta
although we observed with minor differences in the locations of Region and Pearl River Delta Region. These are the locations where
more and less polluted areas. For example, estimated levels of PM1 the highest levels of PM2.5 were estimated in previous studies (Ma
G. Chen et al. / Environmental Pollution 233 (2018) 1086e1094 1091

Fig. 4. Annual mean concentrations of PM1 (mg/m3) in China from 2005 to 2014.

Fig. 5. Mean concentrations of PM1 (mg/m3) in four seasons in China from 2005 to 2014.

et al., 2015; Zheng et al., 2016). This difference could be due to the contributions can be seasonally-dependent (Lee et al., 2006a).
fact the major sources of PM1 and PM2.5 do not necessarily Although, to our knowledge, no national studies on estimating
contribute to the same extent for both size fractions (Cabada et al., PM1 in other countries have been reported, the predicted PM1 level
2004; Vecchi et al., 2004). For example, combustion process of China in this study was much higher than those reported by
including biomass burning can make a relatively greater contri- some regional studies in western countries (Pe rez et al., 2008, 2010;
bution to ambient PM1 than PM2.5 (Perrone et al., 2013), and the Viana et al., 2003). The often severe particulate matter air pollution
1092 G. Chen et al. / Environmental Pollution 233 (2018) 1086e1094

Fig. 6. Trends of PM1 concentrations (mg/m3) in heavily polluted regions and the entire China from 2005 to 2014.

in China is mainly caused by coal combustion, traffic and industrial et al., 2000). Further improvements may also be found by
emissions which are associated with the rapid economic develop- including the outputs from chemical transport models (CTMs) in
ment and expansion of the urban population, especially for mega statistical models. The benefit of this has been demonstrated for
cities including Beijing, Shanghai, Guangzhou and Shenzhen (Chan PM10, PM2.5 and NO2 (de Hoogh et al., 2016; Hamm et al., 2015a).
and Yao, 2008; Xu et al., 2013). In this study, the highest estimated With the use of satellite-retrieved AOD data, this study esti-
levels of PM1 occurred in the South-Western Hebei, Chongqing and mated the temporal and spatial variations of ambient PM1 con-
Sichuan. The heavy air pollution in South-Western Hebei could be centrations across China during past decade. We hope it will
originated from dense local population and industries of steel and provide information for policy makers to allocate resources for the
power (Wang et al., 2014). The high levels of PM1 estimated in prevention and control of severe particulate matter air pollution in
Sichuan and Chongqing might be due to the local landscape. The China, especially for some heavy-polluted regions. Moreover, the
basin-like topography is also characterized by low wind speed results of this study have the potential to link with a range of health
which slow down the dilution of airborne pollutants due to data to further explore the adverse health effects of PM1.
frequent temperature inversions (Li et al., 2015). In addition, rapid However, this study has some limitations. We had limited
industrial and economic growth are apparent in the Sichuan Basin. ground monitoring data from 77 stations included in this study and
High levels of PM1 were also present in North-Eastern China during sparse data for Western China especially, including Xinjiang,
winter season, which might be linked with the cold climate where Xizang, Qinghai and Gansu Province. The prediction of PM1 during
local coal-based heating is used for more than 6 months each year 2005e14 in this study is based on an assumption that the rela-
(Ma et al., 2010). Furthermore, this area is highly industrialized part tionship between PM1 and its predictor variables remained
of China which contributes to poor air quality (Sun et al., 2010). consistent over this time. However, this assumption cannot be
The satellite-retrieved AOD was used to demonstrate the po- verified in China due to unavailability of ground monitoring data
tential to predict recent and historical levels of PM1. Other studies prior to 2013. Also, although DT and DB AOD data were downloaded
have reported the predicted PM2.5 levels across China using MODIS separately and combined to fill in the missing values, a high pro-
AOD data with CV R2 values ranging from 73% to 82% (Fang et al., portion of missing AOD values still existed which limits the ability
2016; Ma et al., 2015; You et al., 2016; Zhang et al., 2016). to detect the daily patterns of PM1 concentrations in China. Finally,
Although the predictive ability of our PM1 models did not exceed to improve predictive ability, we included province as a fixed-effect
those studies, it could plausibly be improved by greater numbers of term in the models for prediction. The disadvantage of this
PM1 monitors, similar to the PM2.5 monitoring network (Hamm approach is that it produces discontinuity at boundaries in some
et al., 2016). Apart from ground-level measurements of air pollut- provinces. PM1, with its smaller particle size than PM2.5, might be
ants, the predictive ability could be improved by adding traffic and more harmful on human health than PM2.5 (Lin et al., 2016).
road information in the model, considering traffic emissions are Considering the importance of PM1 and its potential strong asso-
main sources of outdoor air pollution (Hamra et al., 2015; Künzli ciations with health in China, more exposure data should be
G. Chen et al. / Environmental Pollution 233 (2018) 1086e1094 1093

obtained and future studies should further explore its spatial and scattering for the determination of particulate size distribution, and particulate
mass, including the semi-volatile fraction. J. Air & Waste Manag. Assoc. 59,
temporal distributions.
Guo, J., Xia, F., Zhang, Y., Liu, H., Li, J., Lou, M., He, J., Yan, Y., Wang, F., Min, M., 2017.
5. Conclusion Impact of diurnal variability and meteorological factors on the PM 2.5-AOD
relationship: implications for PM 2.5 remote sensing. Environ. Pollut. 221,
Statistical models using satellite-retrieved AOD, land use infor- Guo, J., Zhang, X., Che, H., Gong, S., An, X., Cao, C.-X., Guang, J., Zhang, H., Wang, Y.-
mation and meteorology could capture the spatial and temporal Q., Zhang, X.-C., 2009. Correlation between PM concentrations and aerosol
variability in ground-level PM1 concentrations in China. This is the optical depth in eastern China. Atmos. Environ. 43, 5876e5886.
Hamm, N.A.S., Finley, A.O., Schaap, M., Stein, A., 2015a. A spatially varying coeffi-
first study to estimate historical levels of PM1 with satellite remote cient model for mapping PM10 air quality at the European scale. Atmos. En-
sensing. It provides important quantitative information regarding viron. 102, 393e405.
the distribution of PM1 across China. The results have the potential Hamm, N.A.S., Soares Magalhaes, R.J., Clements, A.C., 2015b. Earth observation,
spatial data quality, and neglected tropical diseases. PLoS Negl. Trop. Dis. 9,
to link with a wide range of health data and help understand health e0004164.
outcomes in a high pollution country. Given greater ground-based Hamm, N.A.S., van Lochem, M., Hoek, G., Otjes, R., van der Sterren, S., Verhoeven, H.,
measurements of PM1 as well as environmental data and the 2016. In: Close, J.-P. (Ed.), “The Invisible Made Visible”: Science and Technology.
AiREAS: Sustainocracy for a Healthy City. Springer, pp. 51e77.
output of chemical transport models (CTMs), the predictive ability Hamra, G.B., Laden, F., Cohen, A.J., Raaschou-Nielsen, O., Brauer, M., Loomis, D., 2015.
of our models could be extended and improved. Lung cancer and exposure to nitrogen dioxide and traffic: a systematic review
and meta-analysis. Environ. Health Perspect. 123, 1107e1112.
Hu, Waller, L.A., Lyapustin, A., Wang, Y., Liu, Y., 2014a. Improving satellite-driven PM
Competing financial interests less than 2.5 micrometer models with moderate resolution imaging spectror-
adiometer fire counts in the southeastern US.
The authors declare they have no actual or potential competing Hu, X., Waller, L.A., Lyapustin, A., Wang, Y., Liu, Y., 2014b. 10-year spatial and
temporal trends of PM2.5 concentrations in the southeastern US estimated
financial interests.
using high-resolution satellite data. Atmos. Chem. Phys. 14, 6301e6314.
Hu, X., Waller, L.A., Lyapustin, A., Wang, Y., Liu, Y., 2014c. Improving satellite-driven
Acknowledgements PM2. 5 models with Moderate Resolution Imaging Spectroradiometer fire
counts in the southeastern US. J. Geophys. Res. Atmos. 119.
Jing-Mei, Y., Jin-Huan, Q., Yan-Liang, Z., 2010. Validation of aerosol optical depth
Y.G. was supported by Career Development Fellowship of from Terra and Aqua MODIS retrievals over a tropical coastal site in China.
Australian National Health and Medical Research Council (NHMRC, Atmos. Ocean. Sci. Lett. 3, 36e39.
APP1107107). S.L. was supported by the NHMRC Early Career Just, A.C., Wright, R.O., Schwartz, J., Coull, B.A., Baccarelli, A.A., Tellez-Rojo, M.M.,
Moody, E., Wang, Y., Lyapustin, A., Kloog, I., 2015. Using high-resolution satellite
Fellowship (APP1109193) and Seed Funding from the NHMRC aerosol optical depth to estimate daily PM2.5 geographical distribution in
Centre of Research Excellence (CRE)eCentre for Air quality and Mexico city. Environ. Sci. Technol. 49, 8576e8584.
health Research and evaluation (APP1030259). G.C. was supported Kahn, R., Banerjee, P., McDonald, D., 2001. Sensitivity of multiangle imaging to
natural mixtures of aerosols over ocean. J. Geophys. Res. Atmos. 106,
by China Scholarship Council (CSC). 18219e18238.
Kan, H., Chen, B., Hong, C., 2009. Health impact of outdoor air pollution in China:
Appendix A. Supplementary data current knowledge and future research needs. Environ. Health Perspect. 117,
Kloog, I., Nordio, F., Coull, B.A., Schwartz, J., 2012. Incorporating local land use
Supplementary data related to this article can be found at regression and satellite aerosol optical depth in a hybrid model of spatiotem- poral PM2.5 exposures in the Mid-Atlantic states. Environ. Sci. Technol. 46,
Knibbs, L.D., Hewson, M.G., Bechle, M.J., Marshall, J.D., Barnett, A.G., 2014. A national
References satellite-based land-use regression model for air pollution exposure assessment
in Australia. Environ. Res. 135, 204e211.
Cabada, J.C., Pandis, S.N., Subramanian, R., Robinson, A.L., Polidori, A., Turpin, B., Koelemeijer, R., Homan, C., Matthijsen, J., 2006. Comparison of spatial and temporal
2004. Estimating the secondary organic aerosol contribution to PM2.5Using the variations of aerosol optical thickness and particulate matter over Europe.
EC tracer method special issue of aerosol science and technologyon findings Atmos. Environ. 40, 5304e5315.
from the fine particulate matter supersites program. Aerosol Sci. Technol. 38, Künzli, N., Kaiser, R., Medina, S., Studnicka, M., Chanel, O., Filliger, P., Herry, M.,
140e155. Horak, F., Puybonnieux-Texier, V., Que nel, P., Schneider, J., Seethaler, R.,
Cao, J., Xu, H., Xu, Q., Chen, B., Kan, H., 2012. Fine particulate matter constituents Vergnaud, J.C., Sommer, H., 2000. Public-health impact of outdoor and traffic-
and cardiopulmonary mortality in a heavily polluted Chinese city. Environ. related air pollution: a European assessment. Lancet 356, 795e801.
Health Perspect. 120, 373. Lee, H., Liu, Y., Coull, B., Schwartz, J., Koutrakis, P., 2011. A novel calibration approach
Chan, C.K., Yao, X., 2008. Air pollution in mega cities in China. Atmos. Environ. 42, of MODIS AOD data to predict PM2. 5 concentrations. Atmos. Chem. Phys. 11,
1e42. 7991e8002.
Chen, G., Li, S., Zhang, Y., Zhang, W., Li, D., Wei, X., He, Y., Bell, M.L., Williams, G., Lee, S., Cheng, Y., Ho, K., Cao, J., Louie, P.-K., Chow, J., Watson, J., 2006a. PM1. 0 and
Marks, G.B., 2017. Effects of ambient PM 1 air pollution on daily emergency PM2. 5 characteristics in the roadside environment of Hong Kong. Aerosol Sci.
hospital visits in China: an epidemiological study. Lancet Planet. Health 1, Technol. 40, 157e165.
e221ee229. Lee, S.C., Cheng, Y., Ho, K.F., Cao, J.J., Louie, P.K.K., Chow, J.C., Watson, J.G., 2006b.
de Hoogh, K., Gulliver, J., van Donkelaar, A., Martin, R.V., Marshall, J.D., Bechle, M.J., PM1.0 and PM2.5 characteristics in the roadside environment of Hong Kong.
Cesaroni, G., Pradas, M.C., Dedele, A., Eeftens, M., 2016. Development of West- Aerosol Sci. Technol. 40, 157e165.
European PM 2.5 and NO 2 land use regression models incorporating Levy, R., Mattoo, S., Munchak, L., Remer, L., Sayer, A., Patadia, F., Hsu, N., 2013. The
satellite-derived and chemical transport modelling data. Environ. Res. 151, 1e10. Collection 6 MODIS aerosol products over land and ocean. Atmos. Meas. Tech. 6,
Diner, D.J., Beckert, J.C., Reilly, T.H., Bruegge, C.J., Conel, J.E., Kahn, R.A., 2989e3034.
Martonchik, J.V., Ackerman, T.P., Davies, R., Gerstl, S.A., 1998. Multi-angle Im- Li, Y., Chen, Q.L., Zhao, H.J., Wang, L., Tao, R., 2015. Variations in PM10, PM2.5 and
aging SpectroRadiometer (MISR) instrument description and experiment PM1.0 in an urban area of the sichuan basin and their relation to meteorological
overview. Ieee Trans. Geosci. Remote Sens. 36, 1072e1087. factors. Atmosphere 6, 150e163.
Dockery, D.W., Stone, P.H., 2007. Cardiovascular risks from fine particulate air Lin, H., Tao, J., Du, Y., Liu, T., Qian, Z., Tian, L., Di, Q., Rutherford, S., Guo, L., Zeng, W.,
pollution. N. Engl. J. Med. 356, 511e513. 2016. Particle size and chemical constituents of ambient particulate pollution
Fang, X., Zou, B., Liu, X., Sternberg, T., Zhai, L., 2016. Satellite-based ground PM 2.5 associated with cardiovascular mortality in Guangzhou, China. Environ. Pollut.
estimation using timely structure adaptive modeling. Remote Sens. Environ. 208, 758e766.
186, 152e163. Ma, W.-L., Li, Y.-F., Qi, H., Sun, D.-Z., Liu, L.-Y., Wang, D.-G., 2010. Seasonal variations
Friedl, M.A., Sulla-Menashe, D., Tan, B., Schneider, A., Ramankutty, N., Sibley, A., of sources of polycyclic aromatic hydrocarbons (PAHs) to a northeastern urban
Huang, X., 2010. MODIS Collection 5 global land cover: algorithm refinements city, China. Chemosphere 79, 441e447.
and characterization of new datasets. Remote Sens. Environ. 114, 168e182. Ma, Y., Chen, R., Pan, G., Xu, X., Song, W., Chen, B., Kan, H., 2011. Fine particulate air
Furrer, R., Nychka, D., Sain, S., Nychka, M.D., 2009. Package ‘fields’. R Foundation for pollution and daily mortality in Shenyang, China. Sci. Total Environ. 409,
Statistical Computing. Vienna, Austria. 2473e2477.
packages/fields/fields.pdf. last accessed 22 December 2012. Ma, Z., Hu, X., Sayer, A.M., Levy, R., Zhang, Q., Xue, Y., Tong, S., Bi, J., Huang, L., Liu, Y.,
Grimm, H., Eatough, D.J., 2009. Aerosol measurement: the use of optical light 2015. Satellite-based spatiotemporal trends in PM concentrations: China, 2004-
1094 G. Chen et al. / Environmental Pollution 233 (2018) 1086e1094

2013. Environ. Health Perspect. 124, 184e192. Basque Country (Northern Spain): analysis of a 5-year data record and inter-
Ma, Z.W., Hu, X.F., Huang, L., Bi, J., Liu, Y., 2014. Estimating ground-level PM2.5 in pretation of seasonal variations. Atmos. Environ. 37, 2879e2891.
China using satellite remote sensing. Environ. Sci. Technol. 48, 7436e7444. Wang, J., Christopher, S.A., 2003. Intercomparison between satellite-derived aerosol
Meng, X., Fu, Q., Ma, Z., Chen, L., Zou, B., Zhang, Y., Xue, W., Wang, J., Wang, D., optical thickness and PM2. 5 mass: implications for air quality studies. Geophys.
Kan, H., Liu, Y., 2015. Estimating ground-level PM in a Chinese city by combining Res. Lett. 30.
satellite data, meteorological information and a land use regression model. Wang, L., Wei, Z., Yang, J., Zhang, Y., Zhang, F., Su, J., Meng, C., Zhang, Q., 2014. The
Environ. Pollut. 208, 177e184. 2013 severe haze over southern Hebei, China: model evaluation, source
rez, N., Pey, J., Cusack, M., Reche, C., Querol, X., Alastuey, A., Viana, M., 2010.
Pe apportionment, and policy implications. Atmos. Chem. Phys. 14, 3151e3173.
Variability of particle number, black carbon, and PM10, PM2. 5, and PM1 levels Wang, Y.Q., Zhang, X.Y., Sun, J.Y., Zhang, X.C., Che, H.Z., Li, Y., 2015. Spatial and
and speciation: influence of road traffic emissions on urban air quality. Aerosol temporal variations of the concentrations of PM10, PM2.5 and PM1 in China.
Sci. Technol. 44, 487e499. Atmos. Chem. Phys. 15, 13585e13598.
rez, N., Pey, J., Querol, X., Alastuey, A., Lo
Pe pez, J., Viana, M., 2008. Partitioning of Xu, P., Chen, Y., Ye, X., 2013. Haze, air pollution, and health in China. Lancet 382,
major and trace components in PM 10ePM 2.5ePM 1 at an urban site in 2067.
Southern Europe. Atmos. Environ. 42, 1677e1691. Yang, C., Peng, X., Huang, W., Chen, R., Xu, Z., Chen, B., Kan, H., 2012. A time-
Perrone, M., Becagli, S., Orza, J.G., Vecchi, R., Dinoi, A., Udisti, R., Cabello, M., 2013. stratified case-crossover study of fine particulate matter air pollution and
The impact of long-range-transport on PM1 and PM2. 5 at a Central Mediter- mortality in Guangzhou, China. Int. Arch. Occup. Environ. Health 85, 579e585.
ranean site. Atmos. Environ. 71, 176e186. You, W., Zang, Z., Zhang, L., Li, Y., Wang, W., 2016. Estimating national-scale ground-
Rohde, R.A., Muller, R.A., 2015. Air pollution in China: mapping of concentrations level PM25 concentration in China using geographically weighted regression
and sources. PLoS One 10, e0135749. based on MODIS and MISR AOD. Environ. Sci. Pollut. Res. 23, 8327e8338.
Sayer, A., Hsu, N., Bettenhausen, C., Jeong, M.J., 2013. Validation and uncertainty Zhang, H., Hoff, R.M., Engel-Cox, J.A., 2009. The relation between Moderate Reso-
estimates for MODIS Collection 6 “Deep Blue” aerosol data. J. Geophys. Res. lution Imaging Spectroradiometer (MODIS) aerosol optical depth and PM2. 5
Atmos. 118, 7864e7872. over the United States: a geographical comparison by US Environmental Pro-
Sun, Y., Zhou, Q., Xie, X., Liu, R., 2010. Spatial, sources and risk assessment of heavy tection Agency regions. J. Air & Waste Manag. Assoc. 59, 1358e1369.
metal contamination of urban soils in typical regions of Shenyang, China. Zhang, T., Liu, G., Zhu, Z., Gong, W., Ji, Y., Huang, Y., 2016. Real-time estimation of
J. Hazard Mater 174, 455e462. satellite-derived PM2. 5 based on a semi-physical geographically weighted
Vecchi, R., Marcazzan, G., Valli, G., Ceriani, M., Antoniazzi, C., 2004. The role of at- regression model. Int. J. Environ. Res. Public Health 13, 974.
mospheric dispersion in the seasonal variation of PM1 and PM2. 5 concentra- Zheng, Y., Zhang, Q., Liu, Y., Geng, G., He, K., 2016. Estimating ground-level PM 2.5
tion and composition in the urban area of Milan (Italy). Atmos. Environ. 38, concentrations over three megalopolises in China using satellite-derived
4437e4446. aerosol optical depth measurements. Atmos. Environ. 124, 232e242.
Viana, M., Querol, X., Alastuey, A., Gangoiti, G., Mene ndez, M., 2003. PM levels in the